From patches to WSIs: A systematic review of deep Multiple Instance Learning in computational pathology

Abstract

Clinical decision support systems for pathology, particularly those utilizing computational pathology (CPATH) for whole slide image (WSI) analysis, face significant challenges due to the need for high-quality annotated datasets. Given the vast amount of information contained in WSIs, creating such datasets is often prohibitively expensive and time-consuming. Multiple Instance Learning (MIL) has emerged as a promising alternative, enabling training that relies solely on coarse-grained supervision by the fusion of extensive localized information from large-scale wholes, thereby reducing the dependency on costly pixel-level labeling. As a result, MIL has become a pivotal technique in CPATH, driving a surge in related research, particularly over the past five years. This expanding body of work has catalyzed technological innovation, introduced transformative advancements in the field, and been further accelerated by progress in deep learning architectures, large-scale pretraining strategies, and Large Language Models (LLMs). This paper provides a systematic review of recent developments in deep MIL methods, analyzing technological advancements from multiple perspectives, including encoder backbone architectures, encoder pretraining strategies, and MIL aggregation techniques. We present a comprehensive overview of progress in each domain, catalog specific application scenarios, and highlight pivotal contributions that have shaped the field. Finally, we explore emerging research directions and potential future challenges for MIL-based CPATH.